nb_kernel334_x86_64_sse.s

来自「最著名最快的分子模拟软件」· S 代码 · 共 2,248 行 · 第 1/5 页

    movaps %xmm1,nb334_fiyH1(%rsp)
    movaps %xmm2,nb334_fizH1(%rsp)
    movaps %xmm3,nb334_fixH2(%rsp)
    movaps %xmm4,nb334_fiyH2(%rsp)
    movaps %xmm5,nb334_fizH2(%rsp)
    movaps %xmm6,nb334_fixM(%rsp)
    movaps %xmm7,nb334_fiyM(%rsp)
    movaps %xmm8,nb334_fizM(%rsp)

    ## xmm14 = fMx
    ## xmm15 = fMy
    ## xmm13 = fMz
    movaps %xmm14,%xmm0
    unpcklps %xmm15,%xmm14
    unpckhps %xmm15,%xmm0

    addps  %xmm14,%xmm11
    addps  %xmm0,%xmm12

    movhlps  %xmm13,%xmm14 ## fMzc fMzd

    movlps %xmm11,36(%rdi,%rax,4)
    movhps %xmm11,36(%rdi,%rbx,4)
    movlps %xmm12,36(%rdi,%rcx,4)
    movhps %xmm12,36(%rdi,%rdx,4)
    movss  %xmm13,44(%rdi,%rax,4)
    movss  %xmm14,44(%rdi,%rcx,4)
    shufps $1,%xmm13,%xmm13
    shufps $1,%xmm14,%xmm14
    movss  %xmm13,44(%rdi,%rbx,4)
    movss  %xmm14,44(%rdi,%rdx,4)

        ## should we do one more iteration? 
        subl $4,nb334_innerk(%rsp)
        jl    _nb_kernel334_x86_64_sse.nb334_single_check
        jmp   _nb_kernel334_x86_64_sse.nb334_unroll_loop
_nb_kernel334_x86_64_sse.nb334_single_check: 
        addl $4,nb334_innerk(%rsp)
        jnz   _nb_kernel334_x86_64_sse.nb334_single_loop
        jmp   _nb_kernel334_x86_64_sse.nb334_updateouterdata
_nb_kernel334_x86_64_sse.nb334_single_loop: 
        movq  nb334_innerjjnr(%rsp),%rdx        ## pointer to jjnr[k] 
        movl  (%rdx),%eax
        addq $4,nb334_innerjjnr(%rsp)

        movq nb334_pos(%rbp),%rsi
        lea  (%rax,%rax,2),%rax

        ## fetch j coordinates
        movlps (%rsi,%rax,4),%xmm3              ##  Ox  Oy  
        movlps 16(%rsi,%rax,4),%xmm4            ## H1y H1z 
        movlps 32(%rsi,%rax,4),%xmm5            ## H2z  Mx 
        movhps 8(%rsi,%rax,4),%xmm3             ##  Ox  Oy  Oz H1x
        movhps 24(%rsi,%rax,4),%xmm4            ## H1y H1z H2x H2y
        movhps 40(%rsi,%rax,4),%xmm5            ## H2z  Mx  My  Mz
        ## transpose
        movaps %xmm4,%xmm0
        movaps %xmm3,%xmm1
        movaps %xmm4,%xmm2
        movaps %xmm3,%xmm6
        shufps $18,%xmm5,%xmm4 ## (00010010)  h2x - Mx  - 
        shufps $193,%xmm0,%xmm3 ## (11000001)  Oy  - H1y - 
        shufps $35,%xmm5,%xmm2 ## (00100011) H2y - My  - 
        shufps $18,%xmm0,%xmm1 ## (00010010)  Oz  - H1z - 
        ##  xmm6: Ox - - H1x   xmm5: H2z - - Mz 
        shufps $140,%xmm4,%xmm6 ## (10001100) Ox H1x H2x Mx 
        shufps $136,%xmm2,%xmm3 ## (10001000) Oy H1y H2y My 
        shufps $200,%xmm5,%xmm1 ## (11001000) Oz H1z H2z Mz

        ## store all j coordinates in jO  
        movaps %xmm6,nb334_jxO(%rsp)
        movaps %xmm3,nb334_jyO(%rsp)
        movaps %xmm1,nb334_jzO(%rsp)

    movaps %xmm6,%xmm0  ## jxO
    movaps %xmm1,%xmm2  ## jzO
    movaps %xmm3,%xmm1  ## jyO
    movaps %xmm3,%xmm4  ## jyO
    movaps %xmm6,%xmm3  ## jxO
    movaps %xmm2,%xmm5  ## jzO

        ## do O and H1 in parallel
        subps  nb334_ixO(%rsp),%xmm0
        subps  nb334_iyO(%rsp),%xmm1
        subps  nb334_izO(%rsp),%xmm2
        subps  nb334_ixH1(%rsp),%xmm3
        subps  nb334_iyH1(%rsp),%xmm4
        subps  nb334_izH1(%rsp),%xmm5

        movaps %xmm0,nb334_dxOO(%rsp)
        movaps %xmm1,nb334_dyOO(%rsp)
        movaps %xmm2,nb334_dzOO(%rsp)
        movaps %xmm3,nb334_dxH1H1(%rsp)
        movaps %xmm4,nb334_dyH1H1(%rsp)
        movaps %xmm5,nb334_dzH1H1(%rsp)

        mulps %xmm0,%xmm0
        mulps %xmm1,%xmm1
        mulps %xmm2,%xmm2
        addps %xmm1,%xmm0
        addps %xmm2,%xmm0       ## have rsq in xmm0 
        mulps %xmm3,%xmm3
        mulps %xmm4,%xmm4
        mulps %xmm5,%xmm5
        addps %xmm3,%xmm4
        addps %xmm5,%xmm4       ## have rsq in xmm4
        movaps %xmm0,nb334_rsqOO(%rsp)
        movaps %xmm4,nb334_rsqH1H1(%rsp)

        ## do 1/sqrt(x) for O and  H1
        rsqrtps %xmm0,%xmm1
        rsqrtps %xmm4,%xmm5
        movaps  %xmm1,%xmm2
        movaps  %xmm5,%xmm6
        mulps   %xmm1,%xmm1
        mulps   %xmm5,%xmm5
        movaps  nb334_three(%rsp),%xmm3
        movaps  %xmm3,%xmm7
        mulps   %xmm0,%xmm1
        mulps   %xmm4,%xmm5
        subps   %xmm1,%xmm3
        subps   %xmm5,%xmm7
        mulps   %xmm2,%xmm3
        mulps   %xmm6,%xmm7
        mulps   nb334_half(%rsp),%xmm3   ## rinv O - j water 
        mulps   nb334_half(%rsp),%xmm7   ## rinv H1 - j water  

        movaps %xmm3,nb334_rinvOO(%rsp)
        movaps %xmm7,nb334_rinvH1H1(%rsp)

        movq nb334_VFtab(%rbp),%rsi

        ## do O table LJ interaction
        movaps %xmm3,%xmm0
        movaps %xmm0,%xmm1
        mulss  nb334_rsqOO(%rsp),%xmm1   ## xmm1=r 
        mulss  nb334_tsc(%rsp),%xmm1

        cvttps2pi %xmm1,%mm6
        cvtpi2ps %mm6,%xmm3
        subss    %xmm3,%xmm1    ## xmm1=eps 
        movaps %xmm1,%xmm2
        mulss  %xmm2,%xmm2      ## xmm2=eps2 
        pslld   $2,%mm6

        movd %mm6,%ebx
        lea  (%rbx,%rbx,2),%rbx

        ## load dispersion table data into xmm4
        movlps 16(%rsi,%rbx,4),%xmm4
        movlps 24(%rsi,%rbx,4),%xmm6
        movaps %xmm4,%xmm5
        movaps %xmm6,%xmm7
        shufps $0x1,%xmm5,%xmm5
        shufps $0x1,%xmm7,%xmm7
        ## dispersion table YFGH ready in xmm4-xmm7
        mulss  %xmm1,%xmm6      ## xmm6=Geps 
        mulss  %xmm2,%xmm7      ## xmm7=Heps2 
        addss  %xmm6,%xmm5
        addss  %xmm7,%xmm5      ## xmm5=Fp 
        mulss  nb334_two(%rsp),%xmm7            ## two*Heps2 
        addss  %xmm6,%xmm7
        addss  %xmm5,%xmm7 ## xmm7=FF 
        mulss  %xmm1,%xmm5 ## xmm5=eps*Fp 
        addss  %xmm4,%xmm5 ## xmm5=VV 

        movaps nb334_c6(%rsp),%xmm4
        mulss  %xmm4,%xmm7      ## fijD 
        mulss  %xmm4,%xmm5      ## Vvdw6 

        ## save scalar force in xmm3. Update Vvdwtot directly 
        addss  nb334_Vvdwtot(%rsp),%xmm5
        movaps %xmm7,%xmm3 ## fscal 
        movss %xmm5,nb334_Vvdwtot(%rsp)

        ## load repulsion table data into xmm4
        movlps 32(%rsi,%rbx,4),%xmm4
        movlps 40(%rsi,%rbx,4),%xmm6
        movaps %xmm4,%xmm5
        movaps %xmm6,%xmm7
        shufps $0x1,%xmm5,%xmm5
        shufps $0x1,%xmm7,%xmm7
        ## repulsion table YFGH ready in xmm4-xmm7

        mulss  %xmm1,%xmm6      ## xmm6=Geps 
        mulss  %xmm2,%xmm7      ## xmm7=Heps2 
        addss  %xmm6,%xmm5
        addss  %xmm7,%xmm5      ## xmm5=Fp 
        mulss  nb334_two(%rsp),%xmm7            ## two*Heps2 
        addss  %xmm6,%xmm7
        addss  %xmm5,%xmm7 ## xmm7=FF 
        mulss  %xmm1,%xmm5 ## xmm5=eps*Fp 
        addss  %xmm4,%xmm5 ## xmm5=VV 

        movaps nb334_c12(%rsp),%xmm4
        mulss  %xmm4,%xmm7 ## fijR 
        mulss  %xmm4,%xmm5 ## Vvdw12 
        addss  %xmm3,%xmm7

        addss  nb334_Vvdwtot(%rsp),%xmm5
        movss %xmm5,nb334_Vvdwtot(%rsp)

        xorps  %xmm1,%xmm1
        mulss nb334_tsc(%rsp),%xmm7
        mulss %xmm0,%xmm7
        subss  %xmm7,%xmm1

        movaps %xmm1,%xmm0
        movaps %xmm1,%xmm2

        mulss  nb334_dxOO(%rsp),%xmm0
        mulss  nb334_dyOO(%rsp),%xmm1
        mulss  nb334_dzOO(%rsp),%xmm2
        xorps   %xmm3,%xmm3
        xorps   %xmm4,%xmm4
        xorps   %xmm5,%xmm5
        addss   %xmm0,%xmm3
        addss   %xmm1,%xmm4
        addss   %xmm2,%xmm5
        movaps  %xmm3,nb334_fjxO(%rsp)
        movaps  %xmm4,nb334_fjyO(%rsp)
        movaps  %xmm5,nb334_fjzO(%rsp)
        addss   nb334_fixO(%rsp),%xmm0
        addss   nb334_fiyO(%rsp),%xmm1
        addss   nb334_fizO(%rsp),%xmm2
        movss  %xmm0,nb334_fixO(%rsp)
        movss  %xmm1,nb334_fiyO(%rsp)
        movss  %xmm2,nb334_fizO(%rsp)

        ## do  H1 coulomb interaction
        movaps nb334_rinvH1H1(%rsp),%xmm0   ## rinv 
        movaps %xmm0,%xmm1
        mulps  nb334_rsqH1H1(%rsp),%xmm1        ## r
        mulps nb334_tsc(%rsp),%xmm1

        movhlps %xmm1,%xmm2
        cvttps2pi %xmm1,%mm6
        cvttps2pi %xmm2,%mm7    ## mm6/mm7 contain lu indices 
        cvtpi2ps %mm6,%xmm3
        cvtpi2ps %mm7,%xmm2
        movlhps  %xmm2,%xmm3
        subps    %xmm3,%xmm1    ## xmm1=eps 
        movaps %xmm1,%xmm2
        mulps  %xmm2,%xmm2      ## xmm2=eps2 
        pslld   $2,%mm6
        pslld   $2,%mm7

        psrlq $32,%mm6
        movd %mm6,%ebx
        movd %mm7,%ecx
        psrlq $32,%mm7
        movd %mm7,%edx          ## table indices in ebx,ecx,edx 

        lea  (%rbx,%rbx,2),%rbx
        lea  (%rcx,%rcx,2),%rcx
        lea  (%rdx,%rdx,2),%rdx

        movq nb334_VFtab(%rbp),%rsi

        movlps (%rsi,%rbx,4),%xmm4
        movlps (%rsi,%rcx,4),%xmm3
        movlps (%rsi,%rdx,4),%xmm7
        movhps 8(%rsi,%rbx,4),%xmm4
        movhps 8(%rsi,%rcx,4),%xmm3
        movhps 8(%rsi,%rdx,4),%xmm7
        movaps %xmm3,%xmm6
        unpcklps %xmm7,%xmm6
        unpckhps %xmm7,%xmm3
        movaps %xmm4,%xmm5
        movaps %xmm4,%xmm7
        shufps $0x40,%xmm6,%xmm4
        shufps $0xE4,%xmm6,%xmm5
        movaps %xmm7,%xmm6
        shufps $0x48,%xmm3,%xmm6
        shufps $0xEC,%xmm3,%xmm7
        ## coulomb table ready, in xmm4-xmm7

        mulps  %xmm1,%xmm6      ## xmm6=Geps 
        mulps  %xmm2,%xmm7      ## xmm7=Heps2 
        addps  %xmm6,%xmm5
        addps  %xmm7,%xmm5      ## xmm5=Fp 
        mulps  nb334_two(%rsp),%xmm7            ## two*Heps2 

        xorps  %xmm3,%xmm3
        ## fetch charges to xmm3 (temporary) 
        movss   nb334_qqHH(%rsp),%xmm3
        movhps  nb334_qqMH(%rsp),%xmm3
        shufps $193,%xmm3,%xmm3 ## 11000001 

        addps  %xmm6,%xmm7
        addps  %xmm5,%xmm7 ## xmm7=FF 
        mulps  %xmm1,%xmm5 ## xmm5=eps*Fp 
        addps  %xmm4,%xmm5 ## xmm5=VV 
        mulps  %xmm3,%xmm5 ## vcoul=qq*VV  
        mulps  %xmm7,%xmm3 ## fijC=FF*qq 
        ## at this point xmm5 contains vcoul and xmm3 fijC 

        addps  nb334_vctot(%rsp),%xmm5
        movaps %xmm5,nb334_vctot(%rsp)

        mulps  nb334_tsc(%rsp),%xmm3
        xorps  %xmm2,%xmm2
        subps  %xmm3,%xmm2
        mulps  %xmm2,%xmm0
        movaps %xmm0,%xmm1
        movaps %xmm0,%xmm2

        mulps   nb334_dxH1H1(%rsp),%xmm0
        mulps   nb334_dyH1H1(%rsp),%xmm1
        mulps   nb334_dzH1H1(%rsp),%xmm2
        ## update forces H1 - j water 
        movaps  nb334_fjxO(%rsp),%xmm3
        movaps  nb334_fjyO(%rsp),%xmm4
        movaps  nb334_fjzO(%rsp),%xmm5
        addps   %xmm0,%xmm3
        addps   %xmm1,%xmm4
        addps   %xmm2,%xmm5
        movaps  %xmm3,nb334_fjxO(%rsp)
        movaps  %xmm4,nb334_fjyO(%rsp)
        movaps  %xmm5,nb334_fjzO(%rsp)
        addps   nb334_fixH1(%rsp),%xmm0
        addps   nb334_fiyH1(%rsp),%xmm1
        addps   nb334_fizH1(%rsp),%xmm2
        movaps  %xmm0,nb334_fixH1(%rsp)
        movaps  %xmm1,nb334_fiyH1(%rsp)
        movaps  %xmm2,nb334_fizH1(%rsp)

        ## i H2 & M simultaneously first get i particle coords: 
    movaps  nb334_jxO(%rsp),%xmm0
    movaps  nb334_jyO(%rsp),%xmm1
    movaps  nb334_jzO(%rsp),%xmm2
    movaps  %xmm0,%xmm3
    movaps  %xmm1,%xmm4
    movaps  %xmm2,%xmm5
        subps   nb334_ixH2(%rsp),%xmm0
        subps   nb334_iyH2(%rsp),%xmm1
        subps   nb334_izH2(%rsp),%xmm2
        subps   nb334_ixM(%rsp),%xmm3
        subps   nb334_iyM(%rsp),%xmm4
        subps   nb334_izM(%rsp),%xmm5
        movaps %xmm0,nb334_dxH2H2(%rsp)
        movaps %xmm1,nb334_dyH2H2(%rsp)
        movaps %xmm2,nb334_dzH2H2(%rsp)
        movaps %xmm3,nb334_dxMM(%rsp)
        movaps %xmm4,nb334_dyMM(%rsp)
        movaps %xmm5,nb334_dzMM(%rsp)
        mulps %xmm0,%xmm0
        mulps %xmm1,%xmm1
        mulps %xmm2,%xmm2
        mulps %xmm3,%xmm3
        mulps %xmm4,%xmm4
        mulps %xmm5,%xmm5
        addps %xmm1,%xmm0
        addps %xmm3,%xmm4
        addps %xmm2,%xmm0       ## have rsqH2 in xmm0 
        addps %xmm5,%xmm4       ## have rsqM in xmm4 

        ## start with H2, save data 
        movaps %xmm0,nb334_rsqH2H2(%rsp)
        movaps %xmm4,nb334_rsqMM(%rsp)
        ## do invsqrt 
        rsqrtps %xmm0,%xmm1
        rsqrtps %xmm4,%xmm5
        movaps  %xmm1,%xmm2
        movaps  %xmm5,%xmm6
        mulps   %xmm1,%xmm1
        mulps   %xmm5,%xmm5
        movaps  nb334_three(%rsp),%xmm3
        movaps  %xmm3,%xmm7
        mulps   %xmm0,%xmm1
        mulps   %xmm4,%xmm5
        subps   %xmm1,%xmm3
        subps   %xmm5,%xmm7
        mulps   %xmm2,%xmm3
        mulps   %xmm6,%xmm7
        mulps   nb334_half(%rsp),%xmm3   ## rinv H2 - j water 
        mulps   nb334_half(%rsp),%xmm7   ## rinv M - j water  

        movaps %xmm3,nb334_rinvH2H2(%rsp)
        movaps %xmm7,nb334_rinvMM(%rsp)

        movaps %xmm3,%xmm1
        mulps  nb334_rsqH2H2(%rsp),%xmm1        ## xmm1=r 
        movaps %xmm3,%xmm0      ## xmm0=rinv 
        mulps  nb334_tsc(%rsp),%xmm1

        movhlps %xmm1,%xmm2
        cvttps2pi %xmm1,%mm6
        cvttps2pi %xmm2,%mm7    ## mm6/mm7 contain lu indices 
        cvtpi2ps %mm6,%xmm3
        cvtpi2ps %mm7,%xmm2
        movlhps  %xmm2,%xmm3
        subps    %xmm3,%xmm1    ## xmm1=eps 
        movaps %xmm1,%xmm2
        mulps  %xmm2,%xmm2      ## xmm2=eps2 
        pslld   $2,%mm6
        pslld   $2,%mm7

        psrlq $32,%mm6
        movd %mm6,%ebx
        movd %mm7,%ecx
        psrlq $32,%mm7
        movd %mm7,%edx          ## table indices in ebx,ecx,edx 

        lea  (%rbx,%rbx,2),%rbx
        lea  (%rcx,%rcx,2),%rcx
        lea  (%rdx,%rdx,2),%rdx

        movlps (%rsi,%rbx,4),%xmm4
        movlps (%rsi,%rcx,4),%xmm3
        movlps (%rsi,%rdx,4),%xmm7
        movhps 8(%rsi,%rbx,4),%xmm4
        movhps 8(%rsi,%rcx,4),%xmm3
        movhps 8(%rsi,%rdx,4),%xmm7
        movaps %xmm3,%xmm6
        unpcklps %xmm7,%xmm6
        unpckhps %xmm7,%xmm3
        movaps %xmm4,%xmm5
        movaps %xmm4,%xmm7
        shufps $0x40,%xmm6,%xmm4
        shufps $0xE4,%xmm6,%xmm5
        movaps %xmm7,%xmm6
        shufps $0x48,%xmm3,%xmm6
        shufps $0xEC,%xmm3,%xmm7
        ## coulomb table ready, in xmm4-xmm7

        mulps  %xmm1,%xmm6      ## xmm6=Geps 
        mulps  %xmm2,%xmm7      ## xmm7=Heps2 
        addps  %xmm6,%xmm5
        addps  %xmm7,%xmm5      ## xmm5=Fp 
        mulps  nb334_two(%rsp),%xmm7            ## two*Heps2 

        xorps  %xmm3,%xmm3

        ## fetch charges to xmm3 (temporary) 
        movss   nb334_qqHH(%rsp),%xmm3
        movhps  nb334_qqMH(%rsp),%xmm3
        shufps $193,%xmm3,%xmm3 ## 11000001

        addps  %xmm6,%xmm7
        addps  %xmm5,%xmm7 ## xmm7=FF 
        mulps  %xmm1,%xmm5 ## xmm5=eps*Fp 
        addps  %xmm4,%xmm5 ## xmm5=VV 
        mulps  %xmm3,%xmm5 ## vcoul=qq*VV  
        mulps  %xmm7,%xmm3 ## fijC=FF*qq 
        ## at this point xmm5 contains vcoul and xmm3 fijC 
        addps  nb334_vctot(%rsp),%xmm5
        movaps %x